As part of The State of the AI Frontier 2026, AI Frontier Network invited leaders building and deploying AI in the real world to share where the frontier is actually moving. In this contribution, Pavlo Martinovych, Senior Product Manager at Uptiq AI, gives a candid read on what changes in 2026 — and what to watch.
On the capability that will reach the mainstream
Long-memory, long-running agents and the expanded context windows that make them possible. The Agent that keeps making forward progress on a goal across many sessions, many data sources, many systems, and potentially many days or weeks, while leaving the workspace clean enough that the next session can resume where the last one left off.
This is the shift that turns AI from a task automator into something closer to a digital colleague. It is also the shift that makes an intelligence layer possible inside a complex enterprise - one that sits on top of existing systems and holds the rules, the prior decisions, and the in-flight work across the entire organization, instead of fragmenting that knowledge across hundreds of disconnected prompts and point solutions.
On AI and decision-making
Most large enterprises spent the last two years deploying AI fast - under competitive pressure, vendor by vendor, workflow by workflow: a fraud model from one provider, an underwriting model from another, a customer service agent from a third, a personalization engine from a fourth. Decisions got distributed across dozens of systems that are not aware of each other. A decision that used to be a single human moment is now the accumulated output of many smaller calls made in parallel, by systems no one fully sees end to end. By 2027, the harder problem for those organizations is governance, not speed. Can they see what is being decided where, explain it to a regulator or a customer, and intervene when something drifts. Nobody is going to stop and clean it up. Each new model release opens new capabilities, and the competitive pressure pushes everyone to build the next layer on top of the last one. The organizations that handle the next two years well are not the ones that pause to consolidate - they are the ones that build their next AI deployments on top of a policy and audit layer that all the agents share, so the new capabilities compound instead of fragmenting further.
On the most underestimated risk
Security in an agentic architecture. Agents need broad data access to be useful - that is the whole point. So the assumptions enterprise security was built on (least privilege, perimeter defense, anomaly detection on unusual access patterns) start to break. The agent legitimately has access to everything. An attacker no longer needs to escape a sandbox; they need to convince the agent to summarize, format, or transmit data on their behalf. The attack looks identical to legitimate use. Identity is the second layer of the problem. Traditional access control assumes one identity behind one request. Agents act on behalf of users, on behalf of customers, on behalf of other agents - sometimes all at once. Permissions get inherited and chained across systems in ways nobody designed. An agent compromised at one layer can escalate privileges across systems that all trusted it because someone upstream did. This is one of the interesting engineering problems of the next years.
On the advice that matters now
Before you automate something, redesign it. Most companies starting their AI journey are looking for processes to make faster. The faster version of a bad process is still a bad process, just with more leverage behind it. Half the workflows people want to automate exist only because of historical software limitations - a form that has to be filled because a system upstream cannot pass the data, an approval step that exists because someone got burned once a decade ago, a handoff between teams that is there because the org chart used to look different. The right answer for a sixty-step workflow is usually to delete most of it. The companies that get the most out of AI in the next two years will be the ones that rethink which processes should exist at all, and then build lean versions of the ones that survive. The technology is now powerful enough that the harder question is which processes are actually worth keeping.
On the trend worth watching
What excites me is that the limit on what gets built is increasingly about imagination. AI is going to handle more and more of the precise, mathematical, mechanical work, and very well. What stays human is the part that depends on judgment, taste, and the ability to see a problem from outside its own discipline. The people and teams that will do the most interesting work with AI are the ones whose thinking is broad enough to recognize what is now worth attempting. At the same time, the AI bubble can make it feel like the race is already mostly run, but in reality we are still very early in what AI can actually take over end-to-end. Most of the meaningful work has not been done yet. It is a good time to be building
